An Energy Based Model for Incorporating Sequence Priors for Target-Specific Antibody Design

With the growing demand for antibody therapeutics, there is a great need for computational methods to accelerate antibody discovery and optimization. Advances in machine learning on graphs have been leveraged to develop generative models of antibody sequence and structure that condition on specific antigen epitopes. However, the data availability for training models on structure (∼5k antibody binding complexes Schneider et al. [2022]) is dwarfed by the amount of antibody sequence data available (> 550M sequences Olsen et al. [2022]) which have been used to train protein language models useful for antibody generation and optimization. Here we motivate the combination of well-trained antibody sequence models and graph generative models on target structures to enhance their performance for target-conditioned antibody design. First, we present the results of an investigation into the sitewise design performance of popular target-conditioned design models. We show that target-conditioned models may not be incorporating target information into the generation of middle loop residues of the complementarity-determining region of the antibody sequence. Next, we propose an energy-based model framework designed to encourage a model to learn target-specific information by supplementing it with pre-trained marginal-sequence information. We present preliminary results on the development of this model and outline future steps to improve the model framework.

In NeurIPS 2023: Generative AI and Biology Workshop